@maxwshen

wandering and wondering. Research Scientist at Prescient Frontiers Research, Genentech. Forbes 30u30 Science.

 San Francisco, CA
maxwshen.com
Joined July 2012
  • Tweets73
  • Following494
  • Followers1.3K
4Media
Are equivariant architectures worth it for guaranteeing 3D-rotational symmetry when modeling proteins? We study this in *Training Dynamics of Learning 3D-Rotational Equivariance* (Oral at NeurIPS AI4Science '25, accepted at TMLR) 🧵👇
2
5
17
4,032
more replies
Check out our poster and oral presentation at NeurIPS AI4Science, and I'd be happy to chat! I am a researcher in Frontiers Research in Prescient Design, Genentech with @kchonyc and we may be hiring soon!
1
2
4
1,054
arxiv.org/abs/2512.02303

Training Dynamics of Learning 3D-Rotational Equivariance

While data augmentation is widely used to train symmetry-agnostic models, it remains unclear how quickly and effectively they learn to respect symmetries. We investigate this by deriving a...

arxiv.org
153
argmax.blog/posts/deep-retri… The deep retrieval + remix hypothesis

The deep retrieval + remixing hypothesis

How do modern deep generative models produce high-quality outputs?

argmax.blog
1
502
argmax.blog/posts/data-two-v… What is data? The classical and postmodern views

What is data? The classical and postmodern views

Is data sacred?

argmax.blog
6
673
In capitalism, narrow AI may be the greatest threat to AGI. argmax.ghost.io/on-creating-…
2
New blog post - argmax.ghost.io/treating-hig…. A unifying and simple concept - iteratively refining high dimensional output seems to work surprisingly well in practice
2
1
12
Why? Perhaps consider one intuition: compared to predicting the optimal y in a single pass, traversing a learned gradient field is more robust to prediction errors: 50% of directions are descent directions!
1
2
Join us at Genentech in a unique Research Scientist , Machine Learning role: 80% research freedom, 20% applied! Drive your own research agenda, collaborate with academia, and build your own group. Find me at #ICML2022 - very happy to chat! roche.wd3.myworkdayjobs-impl…
51
1
227
I started a fun little machine learning blog! argmax.ghost.io/ Two initial posts exploring some well-known topics ...
1
16
137
1. KL Divergence in Discrete Spaces - the forward KL can be related to the multinomial and dirichlet likelihoods argmax.ghost.io/kl-divergenc…
1
8
2. Implicit differentiation through equilibria: attempting a clearer exposition for my own sake argmax.ghost.io/implicit-dif…
9
I am hiring for a pure ML research scientist role: 80% publishing with complete research freedom, 20% consulting and collaborating on applied ML, in a legendary biotech company with a vibrant scientific environment and academia-like culture. careers.gene.com/us/en/job/G….
8
49
2
258
Humbled to be selected for the Forbes 30 under 30 in Science! forbes.com/30-under-30/2022/…

Forbes 30 Under 30 2022: Science

Inventing the future from the atom up.

forbes.com
6
4
1
86
Load more